Google unveils AI model trained on a trillion minutes of wearable health data

Google Alphabet Inc. ↗ 0.78% Research on Wednesday introduced SensorFM, a foundation model for wearable health that was pretrained on more than one trillion minutes of sensor data from five million people, marking what the company describes as a step toward general-purpose AI for continuous health monitoring.research+1

A Single Model for 35 Health Tasks

Unlike conventional wearable health models built for one outcome at a time, SensorFM learns a single reusable representation of human physiology that transfers across cardiovascular, metabolic, sleep, and mental health domains, as well as lifestyle and demographic factors. The model ingests 34 one-minute aggregate features from five sensor modalities — including heart rate, blood-oxygen saturation, skin temperature, and motion — captured from Fitbit and Pixel Watch devices across more than 100 countries.research

When evaluated across 35 health prediction tasks drawn from three independent clinical studies involving nearly 14,000 participants, frozen SensorFM embeddings with only a lightweight linear head outperformed feature-engineered supervised baselines on 34 of the 35 tasks. The model proved especially capable at detecting conditions like depression and anxiety that leave faint traces in sensor data.alphasignal+1

Scaling Without Saturation

The research team, led by Senior Research Scientist Xin Liu and Staff Research Scientist Daniel McDuff, found that scaling both model size and data volume together produced near-linear performance gains with no sign of saturation. The largest variant, SensorFM-B, trained on the full five-million-person corpus, reduced reconstruction loss by 31 percent over the smallest model and improved downstream classification performance by an average of 9 percent.research

To automate adaptation, the researchers built an agentic "classroom" of collaborating LLM agents that iteratively generate, test, and refine prediction heads on SensorFM embeddings — exploring more than 30,000 candidate solutions across experiments.research

Grounding a Personal Health Agent

In a clinician-evaluated study, integrating SensorFM into a Personal Health Agent produced summaries rated as well as those grounded in actual clinical measurements across all five evaluation dimensions, with no statistically distinguishable difference between model predictions and ground-truth labels. The underlying paper, titled "Towards a General Intelligence and Interface for Wearable Health Data," was first posted on arXiv in May and formally announced via the Google Research blog on July 9.arxiv+1